Tencent Hunyuan Releases Hy4: First to Support Vision and Computer Use, Weights Planned to Be Open-Sourced

Tencent Hunyuan officially releases Hy4: for the first time it can see images and operate a computer, open-source weights are still on the way, and the team is very candid about the gap between Hy4 and top closed-source models!

Hy4 adds vision and Computer Use for the first time, targeting finance, office work, front-end games and scientific research. It has entered the strongest open-source camp, but the team also candidly admits there is still a clear gap with top closed-source models; the weights are "planned to be open-sourced," and for now it is mainly available through the API and Tencent products.

https://t.co/yU0yMEtDcA

# Model information

Hy4 has about 770B total parameters, a MoE architecture, and a 1M token context. Compared with the Hy4 preview in August, the official version greatly increased pre-training and post-training. There are two new capabilities: for the first time it can see images, performing fine-grained recognition, complex chart analysis and visual mathematical reasoning; and for the first time it supports Computer Use, operating both the command line and graphical interfaces, and it has also been specifically trained on professional software such as Blender and CAD, as well as long-cycle cross-application tasks.

# Several flagship scenarios

In finance, Tencent and FiT built several thousand workspaces and tens of thousands of tasks, covering fund research, individual stock analysis, quant, asset allocation, corporate finance and tax. 213 financial practitioners participated in the evaluation, of whom 153 came from external institutions and more than 70% had over five years of experience, for a total of 579 real tasks. On a four-point scale, Hy4 scored 3.04, DeepSeek-V4.1-Flash-1017 scored 2.97, and both Kimi K3.1 Pro and GLM-5.3 scored 2.68. The hallucination rate on search-related tasks dropped from 15.76% in the preview version to 5.45%, and the proportion of errors caused by guessing dropped from 23.1% to 5.6%. Collaborative editing of Word, Excel and PowerPoint was also strengthened; partner Manus says the office document task completion rate is close to 97%, which is a figure the partner reported itself.

In office and collaboration, it averaged 83.18 in the WorkBuddy evaluation, with a task pass rate of about 96%. After connecting to WeCom, it can link chats, emails, documents, schedules and meetings to organize multi-project progress, to-dos and priorities.

In front-end and games, it strengthened complex interaction, business logic and state management, and deepened adaptation to UE, Unity and Godot. Compared with the preview version, GameDev and GameLogic improved by about 19.1% and 9.8% respectively; this is compared with itself, not with others.

In scientific research, the internal benchmark TenureTrail has seven disciplines and 123 executable research tasks. Hy4 ranks first among open-source weight models and roughly third among all compared models (about 53.6, versus about 59 for Claude). It performs relatively well on mathematical construction, physical modeling, earth science, systems engineering and socioeconomic decision-making. One official case: given 84 unordered, partially mirrored 2D slices containing about 6.99 million cells, Hy4 completed a 3D reconstruction of a mouse embryo in about 1.5 hours offline, without specialized registration tools.

# AI R&D and open-source by-products

Hunyuan is preparing for "recursive self-improvement." Equipped with the Hyra 2.0 development framework, Hy4 can optimize the pre-training infrastructure of smaller models with the same architecture and automatically train variants that surpass the official Jev baseline on multiple evaluations. It also open-sourced two foundation models: HyCell for single-cell representation and HyPDE for physical field modeling.

# How does it perform on public leaderboards

Tencent ran its own set of comparisons. On non-internal public-style evaluations, Hy4 mostly ranks second to fourth; it has no overall first place in the twelve main tables, and one of its best results is second on E-Bench 2.0. Within the open-source camp it is roughly on par with Kimi K3.1 Pro, with each winning some and losing some: code, 3D modeling, retrieval and scientific research favor Hy4, while terminal automation and scientific terminal tasks favor Kimi.

On the Agent evaluation APEX-Agents v1.1, Hy4 scores 75.2 in Tencent's table, slightly above the Claude Opus 5.5 it measured (about 74.8) and DeepSeek (about 72.9). But on most other rows, Opus still leads. Tencent itself also wrote: there is still a clear gap with top closed-source models, especially in frontier mathematics, ultra-long development tasks, model iteration and visual question answering. For example, Hy-FrontierMath T5 is only about 13, far from the near-perfect scores of the GPT family.

# Pricing, access points and safety

Pricing is the same as the preview version: 6 yuan per million input tokens, 18 yuan for output, and 0.3 yuan for cache hits. Tencent Cloud TokenHub offers a 50% discount until 2026-12-31. WorkBuddy / CodeBuddy have a certain free quota from November 4 to 15. It is already available on TokenHub, OpenRouter, WorkBuddy/CodeBuddy, Yuanbao, WeCom "Dayuan," ima and others; the weights are "planned to be open-sourced" and will come with a system card, but the official version weights have not yet been released. A cheaper Hy4 flash is also said to be coming soon.

On safety, cybersecurity training retains only defensive content, with no exploitation, reverse engineering or penetration; the biosafety self-test Hy-BioExperimentBench is about 30.4/100, and the team says it has not found a clear lowering of the threshold for carrying out biological harm.